Machine Learning Associate ยท MLOps, ML runtimes and AutoML
MLOps strategy on Databricks
What an MLOps strategy covers: reproducibility, versioning of code, data and models, testing, monitoring and automation across dev, staging and prod.
Key points
- Separate dev, staging and prod environments (each with its own Unity Catalog catalog).
- Keep code in Git, data and features in governed Delta tables, and track every run and model with MLflow.
- Test at each stage, promote code through the environments, mark models with aliases such as Champion and Challenger, and monitor for drift with retraining as needed.
Source: https://docs.databricks.com/aws/en/machine-learning/mlops/mlops-workflow
Practice questions
Which tool records parameters, metrics and artifacts for every training run so results can be reproduced and compared?
- MLflow Tracking
- Git folders
- Delta table history
- Unity Catalog lineage
Show answer
Answer: A. MLflow Tracking
MLflow Tracking logs parameters, metrics, code snapshots and artifacts per run. Git versions code, Delta tables hold data, and Unity Catalog governs assets, but none of them is the run-tracking tool.
Which practices belong in an MLOps strategy?
- Separate dev, staging and prod environments
- Unit and integration tests run on each change
- Editing production notebooks directly for quick fixes
- Monitoring data drift and prediction quality
- Overwriting training data in place without versions
Show answer
Answer: A. Separate dev, staging and prod environments, B. Unit and integration tests run on each change, D. Monitoring data drift and prediction quality
Separate environments and automated tests are core practices, and so is monitoring for drift. Letting every user edit production code directly and skipping versioning of data both break reproducibility and control.
True or false: in the recommended setup, data scientists normally have full write and compute access in the production environment.
Show answer
Answer: False
False. Production is tightly controlled and owned by ML engineers; data scientists typically work in dev with read-write access and have at most read-only access to prod data.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
